We propose a Mutual Information (MI)-based framework for balanced classification of Spanish-language tourist reviews using the Rest-Mex 2025 dataset. Our system predicts sentiment polarity, business type (hotel, restaurant, attraction), and geographic location (state and municipality) from over 200,000 annotated entries. We address class imbalance through redundancy pruning and synthetic data augmentation while weighting feature tokens using normalized MI scores. These scores are computed across class labels to capture the discriminative power of each term. Combined with FastText classifiers and rich preprocessing pipelines, our MI-driven approach improves fairness, interpretability, and accuracy in multi-label tourism classification tasks. Results show strong performance in business type classification (F1 = 0.9687) and improved balance across minority classes. This work highlights the potential of combining statistical information-theoretic measures with modern NLP pipelines for real-world tourism sentiment analysis.
Travel & Tourism Competitiveness Report (TTCR) [
Two years later, the outlook of the T&T industry shifted dramatically. The COVID-19 pandemic dealt
a severe blow to travel demand, causing widespread disruption through lockdowns, travel bans, and
the collapse of international mobility. These efects were not limited to businesses but extended to
economies reliant on tourism. Although signs of recovery are now evident, they vary notably across
regions and markets. The path forward is further complicated by global events such as the war in
Ukraine [
These disruptions have likely led to lasting transformations in both the industry and traveler behavior. Tourists are now more attentive to health and sanitation at destinations, and remain wary of potential COVID variants, regulatory changes, and travel interruptions. The pause in global travel has also encouraged reflection on the environmental impact of tourism. In response, both governments and tourism enterprises have begun reevaluating their strategies, reallocating investments, and implementing measures to better manage risk and evolving consumer expectations.
Beyond the pandemic, the tourism sector has undergone a technological shift over the past decade.
Innovations in digitization, information and communication technologies, machine learning, robotics,
and artificial intelligence (AI) have reshaped the way travelers interact with destinations [
Sentiment analysis and classification of user-generated content in tourism ofer significant insights into customer satisfaction, regional trends, and service quality. However, data imbalance and noisy labels present major challenges. While recent NLP pipelines leverage embeddings and deep learning, they often underperform for low-resource classes. This paper presents a hybrid approach that combines Mutual Information (MI)-based feature weighting with class balancing to enhance classification robustness on the Rest-Mex 2025 corpus.
Unlike the past editions [
• Geographic Location: State and municipality in Mexico.
The dataset is multilingual, domain-specific, and imbalanced, with towns like Tulum overrepresented.
Each review is normalized using a custom class, including: • Lowercasing, whitespace trimming. • Stopword removal (extended Spanish list). • Digit normalization via semantic character replacement. • Tokenization (NLTK), Lemmatization (spaCy).
The main contribution of this stage, compared to previous work such as [17], lies in how we extract informative features for each classification subtask (i.e., sentiment polarity, business type, and geographic location). Inspired by earlier applications of Mutual Information (MI) in NLP [18, 19], we leverage MI to quantify the association between words and class labels across the Rest-Mex 2025 dataset.
Mutual Information measures the shared information between two variables, and , and is defined as: (, ) = (, ) log ︂(
(, ) )︂ () ( ) (1) independent, (, ) ≈ classes [20].
where (, ) denotes the joint probability of observing word in class . When and are 0, implying that carries no useful signal for predicting . In contrast, high MI values indicate strong association, making such words discriminative for their corresponding each class ∈ (e.g., sentiment levels or location labels):
In our application, each token ∈ (where is the vocabulary of the corpus) is evaluated against power.
association. • If word occurs uniformly across all classes, (, ) ≈ 0, ofering little to no discriminative • If is mostly exclusive to class , then (, ) > 0, suggesting strong relevance. • If appears frequently in other classes but rarely in , then (, ) < 0, indicating negative To enhance the representational power of each class, we expand the high-MI words by including up to ifve synonyms retrieved via WordNet, assigning each synonym the same MI score as the original word. This results in a trained feature set Ω for class , where each element is a tuple , = (,, , ), representing a word and its normalized MI score. This enriched feature space captures both statistical relevance and semantic diversity, providing a more robust foundation for classification [21].
For overrepresented classes: remove redundant reviews via hybrid Jaccard/Fuzzy similarity. For underrepresented classes: generate synthetic reviews using MI-rich vocabulary templates. All classes are equalized to the global mean number of instances ¯.
We train three FastText classifiers on balanced, MI-weighted inputs: • Polarity model (5-class ordinal). • Business type model (3-class nominal).
• Location model (40-class multiclass).
Prediction follows softmax scoring, with highest probability label chosen per model.
MI-based features can not improved minority class performance either reduced confusion in adjacent polarities.
In this work, we proposed a Mutual Information-driven approach for balanced text classification in the context of the Rest-Mex 2025 challenge. By integrating MI-based feature selection with strategic data balancing—through redundancy reduction and synthetic augmentation—we aimed to mitigate the limitations caused by class imbalance and noisy inputs. The MI scores helped us identify discriminative terms for each class, improving interpretability and enabling targeted vocabulary expansion via synonym enrichment.
Compared to previous MI-based models, our methodology benefits from a more refined preprocessing pipeline and multi-label FastText classifiers tailored for sentiment, business type, and geographic prediction. The results show low performance on the business classification task and acceptable results in more complex subtasks like location and sentiment polarity. In the context of this task, MI is not enough for solving problems with many data, as it gets confused.
Nonetheless, challenges remain. Some tokens with high MI values were semantically irrelevant or derived from noisy user-generated content (e.g., queretarcdm, metrocdmx). Future work will focus on automatic filtering of such terms, integrating contextual embeddings, and refining semantic augmentation strategies to further improve robustness and fairness across all classes.
We declare that the present manuscript has been written entirely by the authors and that no generative artificial intelligence tools were used in its preparation, drafting, or editing. dation system, sentiment analysis and covid semaphore prediction for mexican tourist texts, Procesamiento del Lenguaje Natural 69 (2022). [14] M. Á. Álvarez-Carmona, Á. Díaz-Pacheco, R. Aranda, A. Y. Rodríguez-González, L. Bustio-Martínez, V. Muñis-Sánchez, A. P. Pastor-López, F. Sánchez-Vega, Overview of rest-mex at iberlef 2023: Research on sentiment analysis task for mexican tourist texts, Procesamiento del Lenguaje Natural 71 (2023). [15] M. Á. Álvarez-Carmona, Á. Díaz-Pacheco, R. Aranda, A. Y. Rodríguez-González, L. Bustio-Martínez, V. Herrera-Semenets, Overview of rest-mex at iberlef 2025: Researching sentiment evaluation in text for mexican magical towns, volume 75, 2025. [16] J. Á. González-Barba, L. Chiruzzo, S. M. Jiménez-Zafra, Overview of IberLEF 2025: Natural Language Processing Challenges for Spanish and other Iberian Languages, in: Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2025), co- located with the 41st Conference of the Spanish Society for Natural Language Processing (SEPLN 2025), CEUR-WS. org, 2025. [17] A. Romero-Cantón, R. Aranda, AngelDiaz-Pacheco, J. P. Ramírez-Silva, Mexican epidemiological semaphore color prediction by means of mutual information features, in: CEUR Workshop Proceedings, Coruña, Spain, 2022. [18] R. Guerrero-Rodriguez, M. Á. Álvarez-Carmona, R. Aranda, A. P. López-Monroy, Studying online travel reviews related to tourist attractions using nlp methods: the case of guanajuato, mexico, Current Issues in Tourism (2021) 1–16. doi:https://doi.org/10.1080/13683500. 2021.2007227. [19] R. D. Hjelm, A. Fedorov, S. Lavoie-Marchildon, K. Grewal, P. Bachman, A. Trischler, Y. Bengio, Learning deep representations by mutual information estimation and maximization, arXiv preprint arXiv:1808.06670 (2018). [20] M. Ravanelli, Y. Bengio, Learning speaker representations with mutual information, arXiv preprint arXiv:1812.00271 (2018). [21] M. Á. Álvarez-Carmona, M. Franco-Salvador, E. Villatoro-Tello, M. Montes-y Gómez, P. Rosso, L. Villaseñor-Pineda, Semantically-informed distance and similarity measures for paraphrase plagiarism identification, Journal of Intelligent & Fuzzy Systems 34 (2018) 2983–2990. doi: 10. 3233/JIFS-169483, publisher: IOS Press.